TealDrift addresses OWASP ASI-02 (Agentic Behavior Manipulation) — the risk that an agent’s behavior is subtly altered through prompt injection, model updates, or configuration changes without triggering explicit policy violations.
Why This Matters
Compliance alignment: NIST AI RMF (MEASURE 2.7 — AI system monitoring), ISO 42001 (A.9.4 — monitoring and measurement), EU AI Act Article 9 (risk management system).
Core Concepts
Statistical Baselines
TealDrift builds a baseline profile of normal agent behavior over a configurable window. It tracks:- Action distribution — What actions the agent performs and how often
- Tool usage patterns — Which tools are called, in what order, at what frequency
- Content characteristics — Average content length, vocabulary diversity, sentiment
- Temporal patterns — When the agent is active, request frequency over time
- Error rates — Normal failure rate vs. current failure rate
Drift Score
Each evaluation produces a drift score (0.0 - 1.0) indicating how far current behavior deviates from the baseline:Class
DriftOptions
evaluate()
Evaluate a request against the agent’s behavioral baseline.DriftResult
min_samples Guard
TealDrift requires a minimum number of observations before activating drift detection. This prevents false positives during the initial learning period.MCP Definition-Drift Monitoring
TealDrift monitors MCP tool definitions for unauthorized changes — new tools added, parameters modified, or tools removed.DefinitionDriftResult
Baseline Management
Integration with TealEngine
Related Documentation
- TealEngine v1.3 API — Engine integration
- TealClassifier — Content classification feeds drift baselines
- NHI Governance — Per-agent drift tracking
- SOC/IR Pipeline — Drift alerts in SIEM
- OWASP Policy Pack — ASI-02 drift detection policies

